It picks the wrong numbers
Ask for "revenue for product X" and the AI finds fifteen similar tables with no hint about which one counts. It picks one — with full confidence.
AI Data Analytics
21% → 95%correct answers — the difference is the foundation, not the AI model
AI analytics that gets the numbers right
Anthropic lets Claude answer 95% of its internal data questions automatically. The breakthrough wasn't a smarter model — it was how the data around it was organised. I build that same foundation for your business, sized for an SMB.
Why most AI analytics projects stall
Connect an AI assistant straight to your business data and it answers confidently — but is only right about one question in five. Not because the AI is weak, but because nobody has told it which numbers are the right ones. That context is the actual product.
AI assistant connected straight to the data21%
Same assistant, with the right data foundation95%
Share of analytics questions answered correctly in Anthropic's internal evaluations, before and after structuring the data context around the model.
The three ways it goes wrong
Its knowledge quietly goes stale
Your business changes: new systems, new definitions, renamed fields. Nobody tells the AI, so accuracy drifts down month by month without anyone noticing.
It can't find what's there
The right answer exists in your systems, but among thousands of fields the AI has no map to it. So it guesses instead of looking it up.
The fix
Four layers that take AI from guessing to knowing
None of this is exotic technology. It's the same discipline Anthropic uses internally — scaled down to what a smaller company actually needs.
95% correct answers
- 01Foundation
One source of truth
We decide which tables and numbers are the official ones. Duplicates and near-copies are retired, and every dataset gets a clear owner.
- 02Definitions
A shared dictionary for your numbers
"Revenue", "active customer", "margin" — each term gets one written definition the AI always checks first. The same question always returns the same answer.
- 03Context
Instructions the AI actually reads
Short guide documents that live next to your data and update when the data changes. This layer is the difference between 21% and 95%.
- 04Validation
Automatic quality checks
Test questions with known answers run automatically on every change. Every answer also shows its source, so you can verify instead of trusting blindly.
Built from the bottom up
Why it can't be a one-off project
An AI analytics setup is never finished — your data keeps changing underneath it. Without upkeep, accuracy drifts down towards 65% within a few months. That's why maintenance is part of the delivery rather than an afterthought: every wrong answer is captured and turned into an improvement.
With ongoing maintenance
95%
Without maintenance
21%
LaunchMonth 1Month 2Month 3
Illustrative curve.
What you get
Data inventory and cleanup plan
A map of what data you have, which tables are duplicates, and who owns what — plus the plan for getting to one source of truth.
Written definitions of your key metrics
Your most important numbers defined once, in plain language, stored where both people and AI read them. Built in standard tooling like dbt or Cube — no lock-in to me.
An AI guide to your data
The instruction library that tells the AI how your business works, with checks that force updates when the data changes.
Automatic quality tests
A test suite with known-correct answers that runs on every change, so accuracy is measured — not assumed.
The assistant itself
Claude connected to your data, ready for questions in plain Swedish or English. Every answer shows which source it used and how fresh the data is.
Upkeep that keeps it honest
A weekly accuracy report and a process that turns every wrong answer into a fix.
How we get there
- 01
Map
Inventory your data, your systems, and the questions you actually want answered.
- 02
Structure
Pick the official datasets and write down the definitions.
- 03
Build
Set up the dictionary, the AI instructions, and the assistant.
- 04
Validate
Test against questions with known answers until accuracy holds.
- 05
Launch
Roll out to the team with a source reference on every answer.
- 06
Maintain
Track accuracy weekly and fold every miss back in as an improvement.
FAQ
- Ideally a modern data warehouse (Snowflake, BigQuery, or similar). But most smaller companies start simpler than that — if your data lives in business systems and spreadsheets today, the first step is getting it into one place, and I help with that too.
- A focused first area — sales reporting, for example — typically takes 3–6 weeks depending on the state of your data today. Covering several parts of the business usually takes 8–16 weeks.
- That's the core of the approach. Automatic checks flag when a change breaks an existing definition, and wrong answers are captured and built back in as fixes. The system is built to be maintained, not rebuilt.
- No. The foundation — one source of truth, shared definitions, AI instructions, and quality tests — works with any AI model. I recommend Claude based on Anthropic's published analytics results, but the same setup works with other models too.
- Almost certainly not. Messy data is the normal starting point — that's why the inventory and cleanup plan comes first. The honest answer after the mapping might be 'fix these three things first', and you get that answer before committing to a build.
What do we need to have in place before starting?+
How long does it take?+
What happens when our data changes?+
Does it have to be Claude?+
Is our data too messy for this?+
Related services